Skip to content
About

The trust layer for AI in the physical world.

Vraelis is building the trust layer for AI in the physical world: defense, critical infrastructure and autonomous systems.

Vraelis by Foremake, Inc

Every machine should be able to prove which AI it runs.

Robots, drones, power grids and defense systems are starting to run AI that changes faster than anyone can audit it. Vraelis is building the record that travels with every model: which AI is in which machine, who approved it, and whether it has changed since.

Built for the companies that make these machines, and for the governments, alliances and international organizations that will oversee them.

Foremake, Inc is the umbrella brand behind Vraelis. Vraelis has two products: Vraelis Prompto, open source and available to anyone, and Vraelis Contour, in private development.

About Foremake, Inc

Why now

AI is leaving the screen and moving into machines.

Built by many hands

One team trains a model, another builds it into a machine, a third runs it in the field.

Hard to see once shipped

After a model leaves the lab, few people can say exactly which version a machine is running.

Mistakes are physical

In the cloud, a wrong answer is a retry. In a machine, a wrong or swapped model is a physical event.

The scale of it

Approvals have to travel with the model, on every continent.

DallasMexico CitySantiagoSan FranciscoWarsawLisbonJohannesburgMunichSingaporeDubaiNairobiBengaluruSydneySeoulJakartaTokyo
How we get there

Lock it. Verify it. Govern it.

Lock

Vraelis Prompto

Record exactly which AI you approved, and know the moment anything changes.

Get Vraelis Prompto
Verify

Vraelis Contour

See which AI each machine is running, side by side with what was approved.

About Vraelis Contour
Govern

The registry for physical AI

One place where AI for the physical world is published, approved and checked, by the companies that build it and the institutions that oversee it.

What we are building
Research

The questions behind the work.

These research areas describe the problems we are exploring. They are not additional shipped products.

  • Model integrity: Whether the model that runs is the model that was approved.
  • Adversarial threats: Inputs, updates and supply paths built to change what a model does.
  • Machine trust: Explicit identity, limited permissions and checks at execution boundaries.
  • Security evidence: Records a person can read, kept in order, including what could not be confirmed.

Our partnerships.

Explore our official relationships with ByteDance, Reddit, Stripe Startups and Cloudflare, with each partnership’s scope and milestones.

Meet our partners